Mistral AI released its Large 4 model to the public on Thursday, making one of Europe’s most capable open-weight systems available for any organisation to download and run on its own machines. The Paris-based startup said the model is designed to help organisations identify and fix cybersecurity vulnerabilities, and that the public version improves on the preview shown three weeks ago after testing with governments and corporate users.
What is being released
Large 4, codenamed Le Chonk, is a mixture-of-experts model with about one trillion total parameters, roughly 52 billion active per token. It supports a one-million-token context window, accepts text and image input, and is fluent in more than 160 languages, including EU languages. Mistral said it trained the model on 3,800 Grace Blackwell processors in European data centres.
The release follows a public preview on 6 October, when Mistral introduced the system. Since then, Mistral said it applied preview feedback while scaling reinforcement learning, producing what it described as major gains in coding, defensive cybersecurity and reasoning efficiency. The public release is therefore the finished version of a system already tested in the open, not a new model.
Availability is broad. The weights are published on Hugging Face under a modified MIT open-weight licence, free to use, modify, fine-tune and commercialize, with one condition: model-as-a-service businesses must first obtain a separate commercial licence from Mistral. The model is also served through Mistral’s own application programming interface and Studio, OpenRouter, its Vibe Code assistant, NVIDIA-hosted endpoints and Microsoft Foundry.
Coding and cyber claims
Mistral’s own figures put Large 4 ahead of other open-weight models on coding. Mistral reported 19.7 percent on Cognition’s private FrontierCode 2.0 benchmark, which tests whether a maintainer would merge an agent’s change, ahead of listed open-weight rivals GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. On Terminal-Bench 4.0 it reported 41 percent, up thirteen points from the preview.
On cybersecurity, Mistral reported a 51 percent score on the Artificial Analysis Cyber Index, which measures discovering vulnerabilities, reproducing and validating them, and patching them without breaking functionality. Mistral said the model handles legitimate defensive tasks without unnecessary refusals, while its guardrails distinguish defensive work from malicious tasks, with a reported 96.6 percent refusal rate on malicious cyber requests.
Those figures are company-reported, and should be read that way. Benchmarks chosen by a model maker tend to favour its strengths, and independent testing will test whether Large 4’s gains hold outside Mistral’s evaluations. Beyond doubt is the release’s scale: a trillion-parameter system any team can run itself differs from a model existing only behind one company’s service.
The dual-use question
Open weights cut both ways, and Mistral said so itself. Vice president of science Pierre Stock said he was uncomfortable living in a world without open weights, adding that once a model is out, access cannot be contained. That protects users from being cut off, but it also means anyone can use the model, including people with hostile intent.
Mistral said it monitored cybersecurity use in live traffic after the preview, observing it refuse black-box attacks against live targets. At the same time, Mistral said it sees potential for the model to uplift attacks, and is modifying its guardrails, with updates promised. Partners such as Orange Cyberdefense receive a penetration-testing and red-teaming variant, subject to vetting and contractual safeguards.
The tension is the same one facing every frontier lab this year, as Google’s own cybersecurity-focused Gemini 4 Carbon, covered in our report on Gemini 4 Carbon, shows. Defensive tools and offensive tools share the same capabilities, and the difference lies in who uses them, on whose systems, and with what permission.
Why It Matters
For European governments, banks and cybersecurity teams, the appeal is control. Unlike leading systems that are mainly accessed through online services, Large 4 can be downloaded and run independently, so sensitive code and data need not leave the organisation’s own systems. That independence also removes dependence on a third-party provider that could change terms, restrict access or withdraw a service.
For the wider open-weight field, Large 4 raises the bar. Mistral called it its most capable openly available cybersecurity model, competing in a field dominated by Chinese models, and reported the highest western open-weight score on the Artificial Analysis Intelligence Index. Whether that lead survives testing matters less than a European lab now shipping at this scale, in the open.
Background
Mistral first previewed Large 4 on 6 October 2026, and later that month teamed up with NVIDIA on engineering AI, as covered in our report on Mistral and NVIDIA engineering AI. Thursday’s release makes the model’s weights publicly downloadable.
Source: Mistral Large 4 release, Large 4 checkpoint details
0 Comments